Learning instruction device and learning instruction method

Abstract

To provide a technique capable of deriving an order of multiple objects to be learned by a learning device.SOLUTION: A learning instruction device 10 for instructing a learning device 12 that learns multiple learning models in order to execute a learning process includes: an acquisition unit that acquires learning processing unit cost of multiple learning models; a derivation unit that derives a learning order of multiple learning models on the basis of the acquired learning processing unit cost; an instruction unit that instructs to the learning device so as to learn according to the derived learning order. The derivation unit is configured to derive the learning order in which the higher the learning processing unit cost of the learning model, the faster the learning model is learned.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a technique for instructing a learning device to learn a learning model. [Background technology]

[0002] Patent Document 1 discloses a model diagnosis device that diagnoses an abnormality in a trained neural network model generated in a vehicle. This model diagnosis device stores output values ​​of the trained neural network model and judges the above based on the results of statistical processing of the output values. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2021-096771 Summary of the Invention [Problem to be solved by the invention]

[0004] However, there are only a limited number of learning devices capable of executing large-scale learning processes, and when making the learning device learn various learning subjects, it is preferable to make the learning proceed in an orderly manner.

[0005] An object of the present invention is to provide a technique for deriving an order of multiple learning objects to be learned by a learning device. [Means for solving the problem]

[0006] In order to solve the above problem, one aspect of the present invention is a learning instruction device that instructs a learning device that learns multiple learning models in sequence to execute a learning process, and includes an acquisition unit that acquires the learning process unit costs of the multiple learning models, a derivation unit that derives a learning order of the multiple learning models based on the acquired multiple learning process unit costs, and an instruction unit that instructs the learning device to learn according to the derived learning order. The derivation unit derives a learning order that gives priority to learning faster for a learning model with a higher learning process unit cost. The acquisition unit acquires the amount of data to be learned by the learning model. The derivation unit derives an hourly cost for the learning process based on the winning bid price and the amount of data for the acquired learning model, and derives a learning order for the multiple learning models based on the derived hourly cost.

[0007] Another aspect of the present invention is a learning instruction method for instructing a learning device that sequentially learns multiple learning models to execute a learning process. This method includes the steps of acquiring a learning process unit price for each of the multiple learning models, deriving a learning order for the multiple learning models based on the acquired multiple learning process unit prices, and instructing the learning device to learn according to the derived learning order. In the deriving step, a learning order is derived that gives priority to learning faster for a learning model with a higher learning process unit price. In the acquiring step, the amount of data to be learned by the learning model is acquired. In the deriving step, the hourly cost of the learning process is derived based on the winning bid price and the amount of data of the acquired learning model, and the learning order of the multiple learning models is derived based on the derived hourly cost. Effect of the Invention

[0008] According to the present invention, it is possible to provide a technique for deriving an order of a plurality of learning objects to be learned by a learning device. [Brief description of the drawings]

[0009] [Figure 1] FIG. 1 is a diagram for explaining an overview of a learning instruction system according to an embodiment; [Diagram 2] FIG. 2 is a diagram illustrating a functional configuration of the learning instruction system according to the embodiment. [Diagram 3] 11 is a flowchart of a learning process according to an embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] 1 is a diagram for explaining an overview of a learning instruction system 1 according to an embodiment. Learning instruction system 1 includes a learning instruction device 10, a learning device 12, and multiple user terminal devices 14a, 14b, and 14c. Learning instruction system 1 accepts requests to use learning device 12 from multiple user terminal devices 14a, 14b, and 14c through auction, and instructs learning device 12 to learn based on the auction results.

[0011] The learning instruction device 10 holds an auction for the use of the learning device 12. A plurality of user terminal devices 14a, 14b, and 14c each hold a learning model to be learned and learning data to be re-learned by the learning model, and wish to execute a process of having the learning model learn the learning data using the learning device 12, and submit bids in the auction for the learning instruction device 10. For example, the first user terminal device 14a submits a bid of 1 million yen, the second user terminal device 14b submits a bid of 700,000 yen, and the third user terminal device 14c submits a bid of 500,000 yen, and the user terminal devices 14a, 14b, and 14c each transmit bid information to the learning instruction device 10.

[0012] When the auction period ends, the learning instruction device 10 derives the auction result based on the bidding information. For example, each of the bids from the user terminal devices 14a, 14b, and 14c is successful. The learning instruction device 10 derives a learning order based on the bid amounts, and instructs the learning device 12 to learn each learning model according to the derived learning order. In other words, the learning instruction device 10 instructs the learning device 12 to execute a learning process so as to learn multiple learning models in order according to the learning order.

[0013] The learning device 12 receives a first learning model and first learning data from the first user terminal device 14a, receives a second learning model and second learning data from the second user terminal device 14b, and receives a third learning model and third learning data from the third user terminal device 14c, and re-learns each learning model.

[0014] In this way, by deriving the learning order based on the bid amounts of the user terminal devices 14a, 14b, and 14c, that is, the learning process unit cost of the learning model, it is possible to derive a valid learning order according to the unit cost.

[0015] 2 shows the functional configuration of the learning instruction system 1 of the embodiment. The various functions of the learning instruction system 1 can be configured in hardware with circuit blocks, memory, and other LSIs, and are realized in software with programs loaded into memory, etc. Therefore, it will be understood by those skilled in the art that the various functions of the learning instruction system 1 can be realized in various forms by hardware alone, software alone, or a combination of both, and are not limited to any one of them.

[0016] The learning instruction device 10 includes an acquisition unit 20, an auction unit 22, a derivation unit 24, an instruction unit 26, and a communication unit 28. The learning device 12 includes a communication unit 30, an execution decision unit 32, and a learning unit 34. The user terminal device 14 includes a storage unit 40, a communication unit 42, an input unit 44, and an output unit 46.

[0017] The user terminal device 14 is a computer owned by a user who wishes to perform learning processing, and may be a mobile terminal. A communication unit 42 of the user terminal device 14 is capable of communicating with the learning instruction device 10 and the learning device 12 via a network. When transmitting information to the learning instruction device 10 and the learning device 12, the communication unit 42 adds identification information of the user terminal device 14.

[0018] The input unit 44 is a touch panel or keyboard that can be operated by the user, and receives bidding information for bidding in the auction. The user transmits the bidding information from the user terminal device 14 to the learning instruction device 10 and waits for the auction result. The bidding information includes the bid amount, the contents of the learning model, the amount of parameters of the learning model, the amount of learning data, etc. The bid amount indicates the unit price of the requested learning process. The contents of the learning model indicate what the learning model will do, and may be, for example, a vehicle traffic congestion amount estimation model, a suspicious person detection model, an endoscopic lesion detection model, etc. The amount of parameters of the learning model indicates the number of parameters that determine the operation of the learning model, and may be an approximate value. The amount of learning data indicates the amount of data to be learned by the learning model.

[0019] The output unit 46 is a display and a speaker that displays the auction results to the user. The storage unit 40 stores the learning model 50, the learning data 52, and the auction information. The auction information includes bidding information and winning bid information indicating that the bid has been won. The learning model 50 and the learning data 52 are transmitted to the learning device 12 when the bid is won.

[0020] The communication unit 28 of the learning instruction device 10 communicates with the learning device 12 and the user terminal device 14 via a network. The acquisition unit 20 acquires bidding information from each of the multiple user terminal devices 14. The acquired multiple pieces of bidding information include the learning processing unit prices of the multiple learning models as bid amounts. The bidding information also includes information indicating the amount of learning data to be trained by the learning model and the contents of the learning model.

[0021] The auction unit 22 accepts bids during the auction period, and when the auction period ends, it determines whether each bid has been successful. The auction unit 22 holds preset winning conditions, and determines whether the contents of the bidding information satisfy the winning conditions. If the contents of the bidding information satisfy the winning conditions, it generates a winning bid number and transmits the winning bid number to the user terminal device 14 corresponding to the successful bid information. The auction unit 22 also transmits the winning bid number to the learning device 12 via the communication unit 42. A winning bid number is not generated for the user terminal device 14 corresponding to the bid information that does not satisfy the winning conditions.

[0022] The derivation unit 24 derives a learning order for a plurality of learning models based on the acquired learning processing unit cost. The learning order may be generated using the winning bid number, or may be information in which the winning bid numbers are ordered. The derivation unit 24 derives a learning order in which the higher the learning processing unit cost of the learning model, the earlier the learning is prioritized. The derivation unit 24 may also derive a learning order in which the winning bid with the highest learning processing unit cost of the learning model is prioritized and earlier the learning is prioritized. When the learning processing unit cost of the first learning model is higher than the learning processing unit cost of the second learning model, the derivation unit 24 may derive a learning order in which the first learning model is prioritized and earlier the learning model than the second learning model. This makes it possible to derive a learning order according to the unit cost.

[0023] When the contents of the learning model satisfy a predetermined priority condition, the derivation unit 24 derives a learning order in which the learning model is given priority for learning early regardless of the learning processing unit cost of the learning model. The predetermined priority condition is a condition that distinguishes the importance of the contents of the learning model, and is satisfied when the contents of the learning model relate to life or health. For example, a model that detects pathology satisfies the priority condition. The predetermined priority condition may be satisfied when information indicating the contents of the learning model contained in the bidding information includes a preset word. The predetermined priority condition is set by an administrator of the learning instruction device 10.

[0024] The derivation unit 24 derives the hourly cost of the process of training the learning model based on the winning bid amount (the winning bid amount is the same as the bid amount) of the acquired learning model and the amount of learning data, and derives the learning order of the multiple learning models based on the derived hourly cost. The time required for the learning process correlates with the amount of learning data, and the derivation unit 24 holds a formula or a two-dimensional map that can derive the hourly cost by inputting the winning bid amount and the amount of learning data. This makes it possible to derive a fair learning order based on the hourly cost of the learning process. The derivation unit 24 may derive the hourly cost of the process of training the learning model based on the winning bid amount, the amount of parameters of the learning model, and the amount of learning data. In other words, the learning process unit cost may be the winning bid amount or may be the hourly cost.

[0025] In either case, the derivation unit 24 sets a learning order for each of the won learning models. If the learning device 12 has any incomplete learning models remaining at the end of the auction, the learning order may be derived by combining the incomplete learning models with the newly won learning model. Also, if the learning device 12 has any learning models remaining that were won in past auctions at the end of the auction, the newly won learning model may be set in the learning order after the learning models won in the past auctions.

[0026] The instruction unit 26 instructs the learning device 12 to learn according to the derived learning order. The instruction unit 26 transmits to the learning device 12 information that associates the successful bid number with the learning order.

[0027] The instruction unit 26 may instruct the learning device 12 to interrupt the learning of the learning model being executed and to learn a learning model that satisfies a predetermined priority condition. In other words, even if the learning device 12 is executing a learning process, the learning process of the learning model that satisfies a predetermined priority condition can be interrupted.

[0028] The communication unit 30 of the learning device 12 receives the successful bid number and information indicating the learning order of the learning model from the learning instruction device 10. The communication unit 30 also receives the successful bid number, the learning model, and the learning data from the user terminal device 14.

[0029] The execution decision unit 32 decides to execute the learning process. The execution decision unit 32 collates the winning bid number transmitted from the learning instruction device 10 with the winning bid number transmitted from the user terminal device 14, and allows the learning model for which the collation is completed to learn. The execution decision unit 32 also decides to execute the learning process according to the learning order acquired from the learning instruction device 10.

[0030] The learning unit 34 executes a learning process. For example, in a vehicle traffic congestion amount estimation model, the learning unit 34 inputs vehicle position information, road information, and the like as learning data, and adjusts the parameters of the learning model. The learning unit 34 does not learn multiple learning models simultaneously, but executes the learning process for each model one by one. When the learning unit 34 completes learning using the learning data, the communication unit 30 transmits the learned learning model to the user terminal device 14.

[0031] 3 is a flowchart of the learning process of the embodiment. First, the learning instruction device 10 starts an auction and starts accepting bids (S10). The timing of the start of the auction may be set irregularly according to the schedule of the learning device 12, or may be set periodically every few months.

[0032] The input unit 44 of the user terminal device 14 receives bid information such as the bid amount and the contents of the learning model input by the user (S12), and the communication unit 42 transmits the bid information to the learning instruction device 10 (S14).

[0033] The acquisition unit 20 of the learning instruction device 10 acquires the bidding information from the learning instruction device 10 (S16) and waits until the auction ends. When the auction ends (S18), the auction unit 22 closes the bidding and determines whether the bidder has won based on the bidding information.

[0034] The derivation unit 24 derives a learning order for the multiple successful learning models based on the learning processing unit price (S20). The derivation unit 24 derives a learning order in which the learning model with the higher learning processing unit price is given priority for learning earlier. The derivation unit 24 issues a winning bid number to the learning model of the successful bid information (S22). The winning bid number is a number that can be identified for each learning model.

[0035] Communications unit 28 transmits a winning bid number indicating that the bid has been made to user terminal device 14 (S24). Storage unit 40 of user terminal device 14 receives and stores the winning bid number (S26). Communications unit 28 transmits the winning bid number indicating that the bid has been made and information indicating the learning order to learning device 12 (S28), and learning device 12 receives and stores the winning bid number (S30).

[0036] The communication unit 42 of the user terminal device 14 transmits the winning bid number, the learning model, and the learning data to the learning device 12 (S32). The execution decision unit 32 of the learning device 12 compares the winning bid number received from the learning instruction device 10 with the winning bid number received from the user terminal device 14 (S34), and permits learning of the learning model for which the comparison has been completed. The learning unit 34 executes a process of having the learning model learn the learning data according to the learning order (S36). When the learning process of the learning unit 34 is completed, the communication unit 30 transmits the learned learning model to the user terminal device 14 (S38).

[0037] It should be understood that the embodiments are merely illustrative, that various modifications are possible in the combination of the components, and that such modifications are also within the scope of the present invention. [Explanation of symbols]

[0038] 1 learning instruction system, 10 learning instruction device, 12 learning device, 14 user terminal device, 20 acquisition unit, 22 auction unit, 24 derivation unit, 26 instruction unit, 28 communication unit, 30 communication unit, 32 execution decision unit, 34 learning unit, 40 memory unit, 42 communication unit, 44 input unit, 46 output unit.

Claims

1. A learning instruction device that instructs a learning device that sequentially learns a plurality of learning models to execute a learning process, an acquisition unit that acquires a learning processing cost per unit of each of a plurality of learning models; A derivation unit that derives a learning order of a plurality of learning models based on the acquired plurality of learning processing unit costs; an instruction unit that instructs the learning device to learn in accordance with the derived learning order; The derivation unit derives a learning order in which a learning model having a higher learning processing unit cost is given priority for learning earlier; The acquisition unit acquires an amount of data to be trained by a learning model, The learning instruction device is characterized in that the derivation unit derives the hourly cost of the learning process based on the winning bid amount and data volume of the acquired learning model, and derives the learning order of multiple learning models based on the derived hourly cost.

2. The acquisition unit acquires information indicating the contents of a learning model, The learning instruction device according to claim 1, characterized in that, when the contents of a learning model satisfy a predetermined priority condition, the derivation unit derives a learning order in which the learning model is given priority for early learning regardless of the learning processing unit cost of the learning model.

3. 3. The learning instruction device according to claim 2, wherein the instruction unit instructs the learning device to interrupt learning of a learning model that is currently being performed and to learn a learning model that satisfies a predetermined priority condition.

4. 1. A learning instruction method for instructing a learning device that sequentially learns a plurality of learning models to execute a learning process, comprising: acquiring a learning processing cost per unit of each of a plurality of learning models; Deriving a learning order of a plurality of learning models based on the acquired plurality of learning processing unit costs; instructing the learning device to learn according to the derived learning sequence; In the step of deriving, a learning order is derived in which a learning model having a higher learning processing unit cost is given priority for learning earlier; In the step of acquiring, an amount of data to be trained by a learning model is acquired; A learning instruction method characterized in that in the derivation step, an hourly cost for the learning process is derived based on the winning bid price and data volume of the acquired learning model, and the learning order of multiple learning models is derived based on the derived hourly cost.

Images